Book a diagnostic
Security and governance

What the agent can do, and what it cannot.

Controls mapped to the risks they actually address. No badges we have not earned, and no claims we cannot show you the evidence for.

Read this first

Neuro is not certified against ISO 27001 or SOC 2. We say so plainly rather than displaying badges we do not hold. What we can show you is the access model, the audit trail and the retention policy for every agent we have shipped, and we will walk your security team through them before any contract.

Risk, then control

Six things that go wrong, and what stops them.

  1. 01
    Risk

    An agent reads records the requester was never allowed to see.

    Identity-scoped retrieval

    The agent inherits the requesting user’s permissions at query time rather than holding a superset of its own. Documents outside that scope are never retrieved, so they cannot appear in an answer.

  2. 02
    Risk

    Confidential documents end up training a third-party model.

    No-training contracts, EU processing

    We only use model providers under agreements that exclude your data from training, and we process inside the European Union. The provider list and its terms are part of the engagement documentation.

  3. 03
    Risk

    An agent changes a customer record and nobody can explain why.

    Full decision trace

    Every run logs the input, the documents retrieved, the tools called and the output. A decision can be replayed months later, which is what makes an audit a walkthrough rather than a rewrite.

  4. 04
    Risk

    A wrong action propagates before anyone notices.

    Tiered writes and approval gates

    Reads are free. Drafts and internal notes apply automatically and are reversible. Anything customer-visible or financial waits for a named human until the test set justifies otherwise.

  5. 05
    Risk

    Conversation history accumulates indefinitely by default.

    Chosen retention windows

    Retention is a design decision made during scoping, not a default. We keep the shortest window the process tolerates, separate prompts from personal data where possible, and treat deletion as a tested operation.

  6. 06
    Risk

    A model change silently degrades quality.

    Regression runs on every change

    The business test set runs on every model, prompt or retrieval change. A drop in score blocks the release. Nothing reaches production on the strength of a demo.

02

Decided before launch.

These four questions are cheaper to answer before a system is connected to real records than after.

GDPR

Processing basis and records

Before an agent touches personal data we document what data it processes, on what legal basis, for how long, and who can read the logs. That belongs in the scoping document, not in a review three weeks before launch.

EU AI Act

Classification and transparency

We classify each agent against the Act’s risk tiers and record the reasoning. Users are told when they are talking to an agent, and every agent has a documented human escalation path.

Access

Identity and revocation

Each agent runs under its own service account with its own scopes, never a borrowed human login. That makes access reviewable and revocation a single step.

Hosting

Where the data sits

Inference and storage happen with providers inside the European Union. The hosting decision is made explicitly at scoping and written into the engagement.

Bring your security team.

We would rather answer hard questions at scoping than at launch. Send your questionnaire and we will work through it.

Book a diagnostic